A mosquito distribution monitoring and early warning method, device and system based on deep learning

Abstract

The application discloses a mosquito distribution monitoring and early warning method, device and system based on deep learning, and relates to the technical field of data processing.The method comprises the following steps: acquiring mosquito density time series data, microclimate time series data, breeding point humidity time series data and global irrigation event records of each monitoring subarea in a target agricultural area; taking each monitoring subarea as a target subarea in turn, performing semantic mining on the breeding point humidity time series data corresponding to the target subarea based on the global irrigation event records, and generating breeding source driving features of the target subarea; performing deep semantic mining on the microclimate time series data and the mosquito density time series data corresponding to the target subarea based on the breeding source driving features, and obtaining breeding outbreak risk features of the target subarea; and generating a mosquito risk heat map based on the breeding outbreak risk features.The application has the effect of improving the mosquito outbreak early warning accuracy caused by irrigation.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, device and system for monitoring and early warning of mosquito distribution based on deep learning. Background Technology

[0002] In large-scale irrigated and greenhouse agriculture, mosquitoes not only harass workers and harm livestock, but also carry and spread pathogens, posing a continuous threat to production safety and crop quality. To achieve cost-effective mosquito control, precise monitoring and early warning of mosquito density and outbreak risk are necessary.

[0003] Common mosquito monitoring and early warning methods largely rely on statistical or traditional machine learning models constructed from historical mosquito density and regional meteorological data (such as temperature, humidity, and rainfall). However, in typical agricultural settings such as farmland and greenhouses with irrigation networks, mosquito breeding exhibits a distinctly anthropogenic characteristic. Valves, joints, and drippers in irrigation networks easily accumulate water periodically and locally due to water pressure changes or residual water discharge after closure. This accumulated water is precisely the primary breeding ground for mosquito larvae hatching and emergence. The humidity changes in these breeding grounds are directly controlled by irrigation schedules, exhibiting a suddenness, localization, and regularity completely different from natural rainfall accumulation. Simultaneously, the microclimate created by the crop canopy affects water evaporation, larval development rates, and adult mosquito survival, further regulating mosquito outbreaks triggered by irrigation water accumulation.

[0004] Current technologies typically treat humidity merely as a general environmental variable, failing to utilize irrigation events as independent breeding source drivers. They also neglect to perform spatiotemporal correlation and deep semantic mining of the local dynamics of breeding site humidity with canopy microclimate and mosquito density. This results in insufficient sensitivity of models to the complete chain of "irrigation—waterlogging—larval hatching—new mosquito emergence—density outbreak," leading to delayed or misjudgments in early warnings and an inability to provide the spatial distribution of mosquito outbreak risks closely related to irrigation activities in different irrigated sub-regions. Therefore, in agricultural scenarios equipped with irrigation networks and irrigation event recording capabilities, there is an urgent need for a monitoring and early warning technology that can deeply integrate data on irrigation events, waterlogging at breeding sites, crop canopy microclimate, and mosquito density to improve the accuracy and spatial resolution of mosquito outbreak warnings. Summary of the Invention

[0005] To improve the accuracy of early warning for mosquito outbreaks caused by irrigation, this application provides a method, device, and system for monitoring and early warning of mosquito distribution based on deep learning.

[0006] Firstly, this application provides a mosquito distribution monitoring and early warning method based on deep learning, employing the following technical solution: A deep learning-based method for mosquito distribution monitoring and early warning includes: Acquire time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events for each monitoring sub-area within the target agricultural area. The mosquito density time-series data is collected by a mosquito trapping device, the microclimate time-series data is collected by a canopy microclimate sensor, and the breeding point humidity time-series data is collected by water accumulation sensors deployed at irrigation network nodes. Each monitoring sub-area is sequentially taken as the target sub-area. Based on the global irrigation event record, semantic mining is performed on the humidity time series data of the breeding points corresponding to the target sub-area to generate the breeding source driving features of the target sub-area. The breeding source driving features characterize the degree of driving effect of water accumulation in the pipe network nodes caused by irrigation activities on larval hatching and new mosquito emergence. Based on the breeding source-driven characteristics, deep semantic mining is performed on the microclimate time-series data and mosquito density time-series data corresponding to the target sub-region to obtain the breeding outbreak risk characteristics of the target sub-region. The breeding outbreak risk characteristics characterize the response pattern of the microclimate environment driven by breeding sources to mosquito density outbreaks. Based on the aforementioned breeding and outbreak risk characteristics, a mosquito risk heat map is generated.

[0007] By employing the aforementioned technical solution, time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events are acquired for each monitoring sub-area within the target agricultural area. Specifically, mosquito density time-series data is collected by mosquito trapping devices, microclimate time-series data by canopy microclimate sensors, and breeding point humidity time-series data by water accumulation sensors deployed at irrigation network nodes. Then, each monitoring sub-area is sequentially designated as a target sub-area. Based on the global irrigation event records, semantic mining is performed on the breeding point humidity time-series data corresponding to the target sub-area. The process involves mining to generate breeding source driving features for the target sub-region. These features characterize the degree to which water accumulation at pipe network nodes caused by irrigation activities drives larval hatching and new mosquito emergence. Based on these breeding source driving features, deep semantic mining is performed on the corresponding microclimate time-series data and mosquito density time-series data to obtain breeding outbreak risk features for the target sub-region. These features characterize the response pattern of the microclimate environment driven by breeding sources to mosquito density outbreaks. Finally, a mosquito risk heatmap is generated based on these breeding outbreak risk features. This invention treats irrigation events as independent breeding source driving signals, integrates multi-source time-series data on breeding point humidity, canopy microclimate, and mosquito density, and performs progressive semantic mining along the complete disaster chain of "irrigation—water accumulation—larval hatching and new mosquito emergence—density outbreak." This accurately captures mosquito density outbreak patterns triggered by irrigation activities and synergistically regulated by microclimate, which helps improve the accuracy of mosquito outbreak early warning in irrigated agriculture scenarios and overcomes the shortcomings of current technologies that neglect irrigation event drivers, leading to delayed response and misjudgments. Simultaneously, the mosquito risk heat map generated based on breeding outbreak risk characteristics can intuitively identify the spatial distribution of risks closely related to irrigation activities in each monitoring sub-area, providing a decision-making basis for precise pesticide application and irrigation management.

[0008] Optionally, the step of performing semantic mining on the time-series humidity data of the breeding points corresponding to the target sub-region based on the global irrigation event records to generate the breeding source driving features of the target sub-region includes: The humidity time-series data of the breeding points are vectorized to obtain a humidity time-series embedding vector; The global irrigation event records are vectorized to obtain irrigation event embedding vectors; The humidity time-series embedding vector and the irrigation event embedding vector are input into the breeding source semantic mining module, wherein the breeding source semantic mining module includes a first gating unit and a second gating unit, and the first gating unit and the second gating unit have weight parameters that are not shared. The first gating unit modulates the humidity temporal embedding vector based on the irrigation event embedding vector to obtain the first irrigation modulation feature. The second gating unit modulates the humidity temporal embedding vector based on the irrigation event embedding vector to obtain the second irrigation modulation feature. The first irrigation modulation feature and the second irrigation modulation feature are fused by mean fusion to obtain the breeding source driving feature of the target sub-region.

[0009] By adopting the above technical solution, in order to generate the breeding source driving features of the target sub-region, the humidity time series data of the breeding points is vectorized to obtain the humidity time series embedding vector. Then, the global irrigation event records are vectorized to obtain the irrigation event embedding vector. The humidity time series embedding vector and the irrigation event embedding vector are then input into the breeding source semantic mining module. The breeding source semantic mining module includes a first gating unit and a second gating unit. The first gating unit and the second gating unit have non-shared weight parameters. Then, through the first gating unit, the humidity time series embedding vector is gated and modulated based on the irrigation event embedding vector to obtain the first irrigation modulation feature. Then, through the second gating unit, the humidity time series embedding vector is gated and modulated based on the irrigation event embedding vector to obtain the second irrigation modulation feature. Finally, the first irrigation modulation feature and the second irrigation modulation feature are fused by mean to obtain the breeding source driving features of the target sub-region.

[0010] Optionally, the step of gating and modulating the humidity temporal embedding vector based on the irrigation event embedding vector to obtain the first irrigation modulation feature includes: A linear mapping is performed on the irrigation event embedding vector to obtain an intermediate mapping vector, wherein the dimension of the intermediate mapping vector is the same as the dimension of the humidity time-series embedding vector; A nonlinear transformation is performed on the intermediate mapping vector to obtain the gated vector; The gate vector is broadcast in the time dimension to generate a gate matrix, wherein the shape of the gate matrix is ​​consistent with that of the humidity time-series embedding vector; The humidity temporal embedding vector is multiplied element-wise with the gating matrix to obtain the first irrigation modulation feature.

[0011] By adopting the above technical solution, in order to generate the first irrigation modulation feature, the irrigation event embedding vector is linearly mapped to obtain an intermediate mapping vector, wherein the dimension of the intermediate mapping vector is the same as the dimension of the humidity time-series embedding vector. Then, the intermediate mapping vector is nonlinearly transformed to obtain a gating vector. Then, the gating vector is broadcast in the time dimension to generate a gating matrix, wherein the gating matrix has the same shape as the humidity time-series embedding vector. Then, the humidity time-series embedding vector and the gating matrix are multiplied element-wise to obtain the first irrigation modulation feature.

[0012] Optionally, the step of performing deep semantic mining on the microclimate time-series data and mosquito density time-series data corresponding to the target sub-region based on the breeding source driving characteristics to obtain the breeding outbreak risk characteristics of the target sub-region includes: The microclimate time-series data and the mosquito density time-series data are vectorized to obtain the microclimate embedding vector and the mosquito density embedding vector, respectively. The breeding source driving features, the microclimate embedding vector, and the mosquito density embedding vector are input into the breeding outbreak risk mining module, wherein the breeding outbreak risk mining module includes a breeding source modulation submodule and an attention submodule; Based on the breeding source modulation submodule, using the breeding source driving feature as the query, attention-weighted modulation is applied to the microclimate embedding vector and the mosquito density embedding vector respectively to obtain the microclimate modulation feature and the mosquito density modulation feature. Based on the attention submodule, using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value, cross-modal attention interaction is performed to obtain attention features; The attention features are globally pooled to obtain the breeding outbreak risk features of the target sub-region.

[0013] By adopting the above technical solution, in order to generate the breeding outbreak risk characteristics of the target sub-region, the microclimate time-series data and mosquito density time-series data are vectorized to obtain microclimate embedding vectors and mosquito density embedding vectors, respectively. Then, the breeding source driving features, microclimate embedding vectors, and mosquito density embedding vectors are input into the breeding outbreak risk mining module, which includes a breeding source modulation sub-module and an attention sub-module. Then, based on the breeding source modulation sub-module, using the breeding source driving features as the query, attention-weighted modulation is performed on the microclimate embedding vectors and mosquito density embedding vectors, respectively, to obtain microclimate modulation features and mosquito density modulation features. Then, based on the attention sub-module, using the mosquito density modulation features as the query and the microclimate modulation features as the key and value, cross-modal attention interaction is performed to obtain attention features. Finally, the attention features are globally pooled to obtain the breeding outbreak risk characteristics of the target sub-region.

[0014] Optionally, the step of performing attention-weighted modulation on the microclimate embedding vector and the mosquito density embedding vector based on the breeding source modulation submodule, using the breeding source driving feature as the query, to obtain the microclimate modulation feature and the mosquito density modulation feature, includes: The breeding source driving features, the microclimate embedding vector, and the mosquito density embedding vector are loaded into the breeding source modulation submodule, wherein the breeding source modulation submodule includes a first attention modulation unit and a second attention modulation unit; Using the first attention modulation unit, the microclimate embedding vector is subjected to attention-weighted modulation with the breeding source driving feature as the query, to obtain the microclimate modulation feature; The second attention modulation unit uses the breeding source driving feature as a query to perform attention-weighted modulation on the mosquito density embedding vector to obtain the mosquito density modulation feature.

[0015] By adopting the above technical solution, in order to generate microclimate modulation features and mosquito density modulation features, the breeding source driving features, microclimate embedding vector, and mosquito density embedding vector are loaded into the breeding source modulation submodule. The breeding source modulation submodule includes a first attention modulation unit and a second attention modulation unit. Then, the first attention modulation unit uses the breeding source driving features as a query to perform attention-weighted modulation on the microclimate embedding vector to obtain the microclimate modulation features. Then, the second attention modulation unit uses the breeding source driving features as a query to perform attention-weighted modulation on the mosquito density embedding vector to obtain the mosquito density modulation features.

[0016] Optionally, the step of using the breeding source-driven features as a query to perform attention-weighted modulation on the microclimate embedding vector to obtain microclimate modulation features includes: A linear transformation is performed on the aforementioned breeding source driving features to generate a modulation query vector; A linear transformation is performed on the microclimate embedding vector to generate a microclimate key vector and a microclimate value vector. The microclimate embedding vector includes embedding vectors for multiple time steps. The microclimate key vector is composed of the key vectors corresponding to each time step, and the microclimate value vector is composed of the value vectors corresponding to each time step. The similarity between the modulation query vector and each of the key vectors is calculated, and the microclimate attention weight distribution is generated by a normalized exponential function. Based on the microclimate attention weight distribution, the value vectors are weighted and summed to obtain the microclimate modulation features.

[0017] By adopting the above technical solution, in order to generate microclimate modulation features, the breeding source driving features are linearly transformed to generate a modulation query vector. Then, the microclimate embedding vector is linearly transformed to generate a microclimate key vector and a microclimate value vector. The microclimate embedding vector includes embedding vectors for multiple time steps. The microclimate key vector is composed of the key vectors corresponding to each time step, and the microclimate value vector is composed of the value vectors corresponding to each time step. Then, the similarity between the modulation query vector and each key vector is calculated, and a microclimate attention weight distribution is generated through a normalized exponential function. Then, based on the microclimate attention weight distribution, each value vector is weighted and summed to obtain the microclimate modulation features.

[0018] Optionally, the step of obtaining attention features by performing cross-modal attention interaction based on the attention submodule, using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value, includes: The mosquito density modulation feature and the microclimate modulation feature are loaded into the attention submodule, wherein the attention submodule includes a first interactive attention unit and a second interactive attention unit, and the first interactive attention unit and the second interactive attention unit have weight parameters that are not shared. The first interactive attention unit performs cross-modal attention interaction using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value to obtain the first interactive attention feature. The second interactive attention unit performs cross-modal attention interaction using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value to obtain the second interactive attention feature. The first interactive attention feature and the second interactive attention feature are fused by mean fusion to obtain the attention feature.

[0019] By adopting the above technical solution, in order to generate attention features, mosquito density modulation features and microclimate modulation features are loaded into the attention submodule. The attention submodule includes a first interactive attention unit and a second interactive attention unit. The first interactive attention unit and the second interactive attention unit have non-shared weight parameters. Then, through the first interactive attention unit, using the mosquito density modulation features as the query and the microclimate modulation features as the key and value, cross-modal attention interaction is performed to obtain the first interactive attention feature. Then, through the second interactive attention unit, using the mosquito density modulation features as the query and the microclimate modulation features as the key and value, cross-modal attention interaction is performed to obtain the second interactive attention feature. Finally, the first interactive attention feature and the second interactive attention feature are fused by mean to obtain the attention feature.

[0020] Optionally, the step of performing cross-modal attention interaction using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value to obtain the first interaction attention feature includes: A linear transformation is performed on the mosquito density modulation features to generate a first interactive query matrix; A linear transformation is performed on the microclimate modulation features to generate a first interaction bond matrix and a first interaction value matrix, respectively. Based on the first interaction query matrix and the first interaction key matrix, calculate the first interaction attention weight matrix; Based on the first interaction attention weight matrix, the first interaction value matrix is ​​weighted and summed to obtain the first interaction attention feature.

[0021] By adopting the above technical solution, in order to generate the first interactive attention feature, the mosquito density modulation feature is linearly transformed to generate the first interactive query matrix. Then, the microclimate modulation feature is linearly transformed to generate the first interactive key matrix and the first interactive value matrix respectively. Then, based on the first interactive query matrix and the first interactive key matrix, the first interactive attention weight matrix is ​​calculated. Then, based on the first interactive attention weight matrix, the first interactive value matrix is ​​weighted and summed to obtain the first interactive attention feature.

[0022] Secondly, this application also provides a mosquito distribution monitoring and early warning device based on deep learning, which adopts the following technical solution: A mosquito distribution monitoring and early warning device based on deep learning includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the method described in any one of the first aspects.

[0023] Thirdly, this application also provides a mosquito distribution monitoring and early warning system based on deep learning, which adopts the following technical solution: A mosquito distribution monitoring and early warning system based on deep learning, comprising: The data acquisition module is used to acquire time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events in each monitoring sub-area within the target agricultural area. The mosquito density time-series data is collected by a mosquito trapping device, the microclimate time-series data is collected by a canopy microclimate sensor, and the breeding point humidity time-series data is collected by water accumulation sensors deployed at irrigation network nodes. The breeding source driving feature generation module is used to take each monitoring sub-area as a target sub-area in sequence, and perform semantic mining on the humidity time series data of the breeding points corresponding to the target sub-area based on the global irrigation event record to generate the breeding source driving feature of the target sub-area. The breeding source driving feature represents the degree of driving effect of water accumulation in the pipe network nodes caused by irrigation activities on larval hatching and new mosquito emergence. The breeding outbreak risk feature mining module is used to perform deep semantic mining on the microclimate time series data and mosquito density time series data corresponding to the target sub-region based on the breeding source driving features, so as to obtain the breeding outbreak risk features of the target sub-region, wherein the breeding outbreak risk features characterize the response pattern of the microclimate environment driven by the breeding source to the mosquito density outbreak; The early warning module is used to generate a mosquito risk heat map based on the breeding and outbreak risk characteristics.

[0024] In summary, this application includes at least the following beneficial technical effects: acquiring time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events for each monitoring sub-area within the target agricultural area. Specifically, the mosquito density time-series data is collected by a mosquito trapping device, the microclimate time-series data is collected by a canopy microclimate sensor, and the breeding point humidity time-series data is collected by water accumulation sensors deployed at irrigation network nodes. Then, each monitoring sub-area is sequentially designated as a target sub-area, and based on the global irrigation event records, the corresponding breeding point humidity time-series data for each target sub-area is analyzed. Semantic mining is performed to generate breeding source driving features for the target sub-region. These features characterize the degree to which water accumulation at pipe network nodes caused by irrigation activities drives larval hatching and new mosquito emergence. Based on these breeding source driving features, deep semantic mining is then performed on the corresponding microclimate time-series data and mosquito density time-series data to obtain breeding outbreak risk features for the target sub-region. These features characterize the response pattern of the microclimate environment driven by breeding sources to mosquito density outbreaks. Finally, a mosquito risk heatmap is generated based on these breeding outbreak risk features. This invention treats irrigation events as independent breeding source driving signals, integrates multi-source time-series data on breeding point humidity, canopy microclimate, and mosquito density, and performs progressive semantic mining along the complete disaster chain of "irrigation—water accumulation—larval hatching and new mosquito emergence—density outbreak." This accurately captures mosquito density outbreak patterns triggered by irrigation activities and synergistically regulated by microclimate, which helps improve the accuracy of mosquito outbreak early warning in irrigated agriculture scenarios and overcomes the shortcomings of current technologies that neglect irrigation event drivers, leading to delayed response and misjudgments. Simultaneously, the mosquito risk heat map generated based on breeding outbreak risk characteristics can intuitively identify the spatial distribution of risks closely related to irrigation activities in each monitoring sub-area, providing a decision-making basis for precise pesticide application and irrigation management. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of an embodiment of this application.

[0026] Figure 2 This is a schematic diagram of the system structure of this application.

[0027] Figure 3 This is a structural block diagram of the computer device described in this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] This application discloses a mosquito distribution monitoring and early warning method based on deep learning.

[0030] Reference Figure 1 A deep learning-based method for monitoring and early warning of mosquito distribution includes: Step S11: Obtain time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events for each monitoring sub-area within the target agricultural area. The mosquito density time-series data is collected by a mosquito trapping device, the microclimate time-series data is collected by a canopy microclimate sensor, and the breeding point humidity time-series data is collected by water accumulation sensors deployed at irrigation network nodes.

[0031] It should be noted that the time-series data on mosquito density, microclimate, breeding site humidity, and global irrigation event records are not collected in isolation, but rather form a multi-source data foundation with causal relationships, built around the mosquito disaster chain of "irrigation-breeding-outbreak." From the perspective of the mosquito life cycle, irrigation activities first create localized water accumulation at pipe network nodes. The time-series humidity data at breeding sites collected by water accumulation sensors directly reflects the availability and stability of mosquito larvae breeding grounds. Global irrigation event records reveal the timing, duration, and spatial extent of the formation of such breeding grounds, serving as the driving force signal for breeding source generation. Based on this, the time-series microclimate data such as temperature, relative humidity, and wind speed collected by canopy microclimate sensors determine the larval development rate, the pupa emergence success rate, and the survival and activity ability of adult mosquitoes, constituting the environmental conditions for the transformation of breeding sources into adult mosquito density. Ultimately, the mosquito trapping device collects adult mosquito numbers at preset time intervals, forming time-series data on mosquito density, reflecting the population response under the combined effects of breeding source driving forces and microclimate conditions. These four types of data correspond to each other in time and space, forming a complete observation chain from driving factors, breeding conditions, transformation environment to outbreak manifestations. In practical applications, the above multi-source data can be spatiotemporally aligned according to monitoring sub-regions, and missing values ​​can be imputed and normalized preprocessed to ensure that subsequent semantic mining can accurately capture the linkage pattern of irrigation activities affecting mosquito density through breeding sources and microclimate environment.

[0032] Step S12: Each monitoring sub-region is taken as the target sub-region in turn. Based on the global irrigation event record, semantic mining is performed on the humidity time series data of the breeding points corresponding to the target sub-region to generate the breeding source driving features of the target sub-region. The breeding source driving features represent the degree of driving effect of water accumulation in the pipe network nodes caused by irrigation activities on larval hatching and new mosquito emergence.

[0033] It should be noted that, in order to accurately depict how irrigation activities ultimately drive larval hatching and new mosquito emergence through water accumulation at pipe network nodes, step S12 performs semantic mining to correlate global irrigation event records with the humidity time-series data at breeding points, generating breeding source driving features for the target sub-region. From the perspective of mosquito breeding processes, the humidity time-series data collected by the water accumulation sensors at irrigation pipe network nodes directly records the dynamic changes of mosquito larvae breeding water bodies from formation, maintenance to disappearance, and is a direct representation of breeding source suitability; while global irrigation event records provide the driving signals for the generation of such water accumulation, namely, information such as irrigation start and end times and irrigation volume. There is a clear causal relationship between the two in time: the occurrence of irrigation events leads to an increase in humidity at pipe network nodes, the existence of water accumulation provides the necessary conditions for egg hatching and larval development, and the disappearance of water accumulation leads to the disappearance of the breeding source. Based on this, in a feasible implementation, humidity time-series data and irrigation event records can be vectorized separately. Then, using a dual-gating unit with non-shared parameters, the humidity time-series embedding vector is gated and modulated with the irrigation event embedding vector as a condition. The gating mechanism can filter and enhance the water accumulation response component in the humidity time series that is highly synchronized with irrigation events, while suppressing background humidity fluctuations unrelated to irrigation. This allows for the extraction of dynamic features of breeding sources that are purely driven by irrigation activities and play a key role in larval hatching and new mosquito emergence. The mean fusion after dual-branch modulation further improves the representational stability of this feature, ensuring that subsequent analysis focuses on irrigation-related breeding source drivers.

[0034] Step S13: Based on the breeding source-driven characteristics, perform deep semantic mining on the microclimate time series data and mosquito density time series data corresponding to the target sub-region to obtain the breeding outbreak risk characteristics of the target sub-region. The breeding outbreak risk characteristics represent the response pattern of the microclimate environment driven by breeding sources to mosquito density outbreaks.

[0035] It should be noted that after obtaining the breeding source driving characteristics, which characterize the degree to which irrigation activities drive larval hatching and new mosquito emergence, step S13 deeply integrates the breeding source driving characteristics with microclimate time-series data and mosquito density time-series data to explore the response pattern of the microclimate environment to mosquito density outbreaks under breeding source driving, and generates breeding outbreak risk characteristics. From the perspective of mosquito population dynamics, breeding source driving characteristics determine the potential replenishment of larvae and newly emerged adult mosquitoes, while canopy microclimate time-series data (temperature, humidity, wind speed, etc.) regulate the actual development rate of these larvae, the success rate of pupa emergence, and the survival and activity range of adult mosquitoes, ultimately manifested as the density time-series changes collected by the mosquito trapping device. Therefore, the relationship between microclimate and mosquito density is not a simple environment-quantity relationship, but a conditional response formed under the background of breeding source driving. Step S13 utilizes the breeding source modulation submodule, using breeding source driving features as queries, to perform attention-weighted modulation on the microclimate embedding vector and mosquito density embedding vector, respectively. This modulated feature highlights microclimate change periods and mosquito density response segments closely related to breeding source conditions. Subsequently, through the cross-modal attention submodule, using mosquito density modulation features as queries and microclimate modulation features as keys and values, interaction can be performed. This captures which microclimate patterns will cause a significant increase in mosquito density under specific breeding source driving intensities. The breeding outbreak risk feature obtained after global pooling of the interactive features integrates the linkage information of the entire chain of "irrigation-driven breeding source - microclimate regulation and transformation - density outbreak response," providing a high semantic level expression for risk identification.

[0036] Step S14: Generate a mosquito risk heat map based on the characteristics of breeding and outbreak risk.

[0037] It should be noted that the breeding outbreak risk characteristics comprehensively express the potential trend of mosquito density outbreaks in each monitoring sub-region under the synergistic effect of breeding source conditions and microclimate environment caused by irrigation activities. This step generates a mosquito risk heat map based on these characteristics, transforming complex risk information into spatially intuitive early warning results. Specifically, the breeding outbreak risk characteristics can be input into a preset risk mapping network, outputting a mosquito outbreak risk score for each sub-region. The higher the score, the greater the likelihood of an abnormal increase in adult mosquito density in the sub-region under the current irrigation event and breeding source conditions. After obtaining the risk scores for all monitoring sub-regions, and combining the geographical location information of each sub-region, a mosquito risk heat map covering the entire target agricultural area is generated through spatial interpolation or heat map rendering methods. This heat map can clearly identify the spatial distribution of high-risk mosquito areas closely related to irrigation activities, allowing plant protection personnel to intuitively identify areas requiring priority intervention, thereby implementing targeted control measures such as precise pesticide application, waterlogging removal, or irrigation system adjustments, improving the efficiency and economy of mosquito control.

[0038] In the above implementation, time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events are acquired for each monitoring sub-area within the target agricultural area. The mosquito density time-series data is collected by a mosquito trapping device, the microclimate time-series data is collected by a canopy microclimate sensor, and the breeding point humidity time-series data is collected by water accumulation sensors deployed at irrigation network nodes. Then, each monitoring sub-area is sequentially designated as a target sub-area. Based on the global irrigation event records, semantic mining is performed on the breeding point humidity time-series data corresponding to the target sub-area. The study generates breeding source driving features for the target sub-region, where the breeding source driving features characterize the degree to which water accumulation in pipe network nodes caused by irrigation activities drives larval hatching and new mosquito emergence. Then, based on the breeding source driving features, deep semantic mining is performed on the microclimate time series data and mosquito density time series data corresponding to the target sub-region to obtain the breeding outbreak risk features of the target sub-region. The breeding outbreak risk features characterize the response pattern of the microclimate environment driven by breeding sources to mosquito density outbreaks. Finally, based on the breeding outbreak risk features, a mosquito risk heat map is generated. This invention treats irrigation events as independent breeding source driving signals, integrates multi-source time-series data on breeding point humidity, canopy microclimate, and mosquito density, and performs progressive semantic mining along the complete disaster chain of "irrigation—water accumulation—larval hatching and new mosquito emergence—density outbreak." This accurately captures mosquito density outbreak patterns triggered by irrigation activities and synergistically regulated by microclimate, which helps improve the accuracy of mosquito outbreak early warning in irrigated agriculture scenarios and overcomes the shortcomings of current technologies that neglect irrigation event drivers, leading to delayed response and misjudgments. Simultaneously, the mosquito risk heat map generated based on breeding outbreak risk characteristics can intuitively identify the spatial distribution of risks closely related to irrigation activities in each monitoring sub-area, providing a decision-making basis for precise pesticide application and irrigation management.

[0039] As a further implementation of the method, the step of semantically mining the time-series humidity data of the breeding points corresponding to the target sub-region based on global irrigation event records to generate breeding source-driven features of the target sub-region includes: Step S21: Vectorize the humidity time series data of the breeding points to obtain the humidity time series embedding vector.

[0040] Step S22: Vectorize the global irrigation event records to obtain the irrigation event embedding vector.

[0041] Step S23: Input the humidity time-series embedding vector and the irrigation event embedding vector into the breeding source semantic mining module. The breeding source semantic mining module includes a first gating unit and a second gating unit. The first gating unit and the second gating unit have non-shared weight parameters.

[0042] Step S24: Through the first gating unit, the humidity time-series embedding vector is gated and modulated based on the irrigation event embedding vector to obtain the first irrigation modulation feature.

[0043] Step S25: Using the second gating unit, the humidity time-series embedding vector is gated and modulated based on the irrigation event embedding vector to obtain the second irrigation modulation feature.

[0044] Step S26: The first irrigation modulation feature and the second irrigation modulation feature are fused by mean fusion to obtain the breeding source driving features of the target sub-region.

[0045] It should be noted that the core idea of ​​this implementation is to use irrigation events as conditional signals and perform conditional semantic filtering on the humidity time-series data of breeding points through a gating mechanism. This allows for the precise extraction of the water accumulation response component triggered by irrigation activities, which plays a practical driving role in larval hatching and new mosquito emergence, from the complex humidity dynamics, thus suppressing the interference of background humidity fluctuations unrelated to irrigation on subsequent risk assessment. The first gating unit and the second gating unit have the same network structure, but use non-shared weight parameters. Both use the irrigation event embedding vector as a condition to perform gating modulation operations on the humidity time-series embedding vector. The difference lies in the fact that the non-shared weights allow the two units to spontaneously learn different emphasis on the correlation patterns between irrigation events and humidity time-series data during training. For example, one unit may focus more on capturing the humidity mutation pattern during the rapid formation of water accumulation after irrigation, while the other unit may focus more on capturing the humidity persistence pattern during the slow receding stage of water accumulation. The two constitute a parallel and complementary feature extraction branch, enriching the modeling granularity of the irrigation-water accumulation response from different perspectives. Although the modulation features extracted by the two units differ in detail, their semantic spaces are perfectly aligned: both are the result of gating and filtering the same humidity temporal embedding vector under the same irrigation event conditions, and their core semantic representation is "water accumulation driven features caused by irrigation activities," with consistent feature dimensions and scales. Therefore, mean fusion of the first and second irrigation modulation features is mathematically and physically feasible. Mean fusion can integrate the complementary information of the two branches, smooth out possible estimation biases in a single branch, improve the robustness and generalization ability of the breeding source driven features, and provide a more stable and reliable input for subsequent breeding outbreak risk mining.

[0046] In the above implementation, in order to generate the breeding source driving features of the target sub-region, the humidity time-series data of the breeding points is vectorized to obtain the humidity time-series embedding vector. Then, the global irrigation event records are vectorized to obtain the irrigation event embedding vector. The humidity time-series embedding vector and the irrigation event embedding vector are then input into the breeding source semantic mining module. The breeding source semantic mining module includes a first gating unit and a second gating unit. The first gating unit and the second gating unit have non-shared weight parameters. Then, through the first gating unit, the humidity time-series embedding vector is gated and modulated based on the irrigation event embedding vector to obtain the first irrigation modulation feature. Then, through the second gating unit, the humidity time-series embedding vector is gated and modulated based on the irrigation event embedding vector to obtain the second irrigation modulation feature. Finally, the first irrigation modulation feature and the second irrigation modulation feature are fused by mean to obtain the breeding source driving features of the target sub-region.

[0047] As a further implementation of the method, the step of gating and modulating the humidity temporal embedding vector based on the irrigation event embedding vector to obtain the first irrigation modulation feature includes: Step S31: Perform a linear mapping on the irrigation event embedding vector to obtain an intermediate mapping vector, wherein the dimension of the intermediate mapping vector is the same as the dimension of the humidity time-series embedding vector.

[0048] Step S32: Perform a nonlinear transformation on the intermediate mapping vector to obtain the gated vector.

[0049] Step S33: Broadcast the gated vector in the time dimension to generate a gated matrix, wherein the shape of the gated matrix is ​​consistent with that of the humidity time-series embedding vector.

[0050] Step S34: Multiply the humidity temporal embedding vector and the gating matrix element by element to obtain the first irrigation modulation feature.

[0051] It should be noted that the technical principle of step S25 is basically the same as that of step S24, and the implementation process of step S25 can be referred to steps S31 to S34.

[0052] In the above implementation, in order to generate the first irrigation modulation feature, the irrigation event embedding vector is linearly mapped to obtain an intermediate mapping vector, wherein the dimension of the intermediate mapping vector is the same as the dimension of the humidity time-series embedding vector. Then, the intermediate mapping vector is nonlinearly transformed to obtain a gating vector. Then, the gating vector is broadcast in the time dimension to generate a gating matrix, wherein the gating matrix has the same shape as the humidity time-series embedding vector. Then, the humidity time-series embedding vector and the gating matrix are multiplied element-wise to obtain the first irrigation modulation feature.

[0053] As a further implementation of the method, the step of performing deep semantic mining on the microclimate time-series data and mosquito density time-series data corresponding to the target sub-region based on the breeding source driving characteristics to obtain the breeding outbreak risk characteristics of the target sub-region includes: Step S41: Vectorize the microclimate time series data and mosquito density time series data respectively to obtain the microclimate embedding vector and the mosquito density embedding vector.

[0054] Step S42: Input the breeding source driving features, microclimate embedding vector and mosquito density embedding vector into the breeding outbreak risk mining module, which includes a breeding source modulation submodule and an attention submodule.

[0055] Step S43: Based on the breeding source modulation submodule, using the breeding source driving feature as the query, attention-weighted modulation is applied to the microclimate embedding vector and the mosquito density embedding vector respectively to obtain the microclimate modulation feature and the mosquito density modulation feature.

[0056] Step S44: Based on the attention submodule, using mosquito density modulation features as the query and microclimate modulation features as the key and value, perform cross-modal attention interaction to obtain attention features.

[0057] Step S45: Perform global pooling on the attention features to obtain the breeding outbreak risk features of the target sub-region.

[0058] It should be noted that the core idea of ​​this implementation method is to use the breeding source driving characteristics as conditional anchors, and to explore the conditional response relationship between the breeding source driving force, microclimate environment, and mosquito density layer by layer downstream along the complete disaster chain of "irrigation - water accumulation - larval hatching and new mosquito emergence - density outbreak", ultimately forming a breeding outbreak risk expression that integrates the complete chain. Specifically, in step S42, the breeding source driving characteristics, microclimate embedding vector, and mosquito density embedding vector are jointly input into the breeding outbreak risk mining module. The basis for this is that the three correspond to the three links of "driving force - condition - response" in the disaster chain: the breeding source driving characteristics represent the potential replenishment intensity of larvae and newly emerged adult mosquitoes, the microclimate embedding vector represents the environmental transformation conditions of density outbreak, and the mosquito density embedding vector reflects the actual response results of the population. Only by jointly modeling the three can we learn what level of density outbreak will be triggered under what breeding source driving intensity and what microclimate conditions, rather than analyzing the simple statistical correlation between environment and density in isolation. Step S43 uses the breeding source driving feature as the query and employs an attention mechanism to weight and modulate the microclimate embedding vector and the mosquito density embedding vector respectively. The aim is to select microclimate change periods and density response segments highly correlated with the current breeding source driving state from the original time-series data, guided by breeding source conditions. This imbues the modulated features with breeding source-conditional semantics, effectively suppressing background noise from irrelevant periods. Building on this, step S44 uses the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value for cross-modal attention interaction. The design intent is to explicitly model the conditional mapping relationship of "how microclimate transforms into mosquito density response under a specific breeding source driving background": the mosquito density modulation feature represents the density response state under breeding source conditions. By retrieving the microclimate modulation feature based on this state, the microclimate pattern matching the density response state can be accurately captured, thereby extracting the contribution pattern of microclimate to mosquito density under breeding source driving. After global pooling, the resulting breeding outbreak risk characteristics integrate the linkage information of the complete disaster chain of "irrigation-driven breeding source - microclimate regulation and transformation - density outbreak response", providing a comprehensive expression with a high semantic level for the subsequent generation of risk heat maps.

[0059] In the above implementation, in order to generate the breeding outbreak risk features of the target sub-region, the microclimate time-series data and mosquito density time-series data are vectorized to obtain microclimate embedding vectors and mosquito density embedding vectors, respectively. Then, the breeding source driving features, microclimate embedding vectors, and mosquito density embedding vectors are input into the breeding outbreak risk mining module. The breeding outbreak risk mining module includes a breeding source modulation sub-module and an attention sub-module. Then, based on the breeding source modulation sub-module, using the breeding source driving features as the query, attention-weighted modulation is performed on the microclimate embedding vectors and mosquito density embedding vectors, respectively, to obtain microclimate modulation features and mosquito density modulation features. Then, based on the attention sub-module, using the mosquito density modulation features as the query and the microclimate modulation features as the key and value, cross-modal attention interaction is performed to obtain attention features. Finally, the attention features are globally pooled to obtain the breeding outbreak risk features of the target sub-region.

[0060] As a further implementation of the method, based on the breeding source modulation submodule, using the breeding source-driven features as queries, the steps of performing attention-weighted modulation on the microclimate embedding vector and the mosquito density embedding vector respectively to obtain the microclimate modulation features and the mosquito density modulation features include: Step S51: The breeding source driving features, microclimate embedding vector and mosquito density embedding vector are loaded into the breeding source modulation submodule, wherein the breeding source modulation submodule includes a first attention modulation unit and a second attention modulation unit.

[0061] Step S52: Using the first attention modulation unit and the breeding source driving feature as the query, the microclimate embedding vector is subjected to attention weighted modulation to obtain the microclimate modulation feature.

[0062] Step S53: Using the second attention modulation unit, the mosquito density embedding vector is subjected to attention-weighted modulation with the breeding source driving feature as the query, to obtain the mosquito density modulation feature.

[0063] In the above embodiments, in order to generate microclimate modulation features and mosquito density modulation features, the breeding source driving features, the microclimate embedding vector, and the mosquito density embedding vector are loaded into the breeding source modulation submodule. The breeding source modulation submodule includes a first attention modulation unit and a second attention modulation unit. Then, the first attention modulation unit performs attention-weighted modulation on the microclimate embedding vector with the breeding source driving features as the query to obtain the microclimate modulation features. Then, the second attention modulation unit performs attention-weighted modulation on the mosquito density embedding vector with the breeding source driving features as the query to obtain the mosquito density modulation features.

[0064] As a further implementation of the method, the step of using breeding source-driven features as a query to perform attention-weighted modulation on the microclimate embedding vector to obtain microclimate modulation features includes: Step S61: Perform a linear transformation on the breeding source driving features to generate a modulation query vector.

[0065] Step S62: Perform a linear transformation on the microclimate embedding vector to generate a microclimate key vector and a microclimate value vector. The microclimate embedding vector includes embedding vectors for multiple time steps. The microclimate key vector is composed of the key vectors corresponding to each time step, and the microclimate value vector is composed of the value vectors corresponding to each time step.

[0066] Step S63: Calculate the similarity between the modulation query vector and each key vector, and generate the microclimate attention weight distribution through a normalized exponential function.

[0067] Step S64: Based on the microclimate attention weight distribution, the value vectors are weighted and summed to obtain the microclimate modulation features.

[0068] In the above implementation, in order to generate microclimate modulation features, a linear transformation is performed on the breeding source driving features to generate a modulation query vector. Then, a linear transformation is performed on the microclimate embedding vector to generate a microclimate key vector and a microclimate value vector. The microclimate embedding vector includes embedding vectors for multiple time steps. The microclimate key vector is composed of the key vectors corresponding to each time step, and the microclimate value vector is composed of the value vectors corresponding to each time step. Then, the similarity between the modulation query vector and each key vector is calculated, and a microclimate attention weight distribution is generated through a normalized exponential function. Then, based on the microclimate attention weight distribution, each value vector is weighted and summed to obtain the microclimate modulation features.

[0069] As a further implementation of the method, the step of obtaining attention features by performing cross-modal attention interaction based on the attention submodule, using mosquito density modulation features as queries and microclimate modulation features as keys and values, includes: Step S71: Load the mosquito density modulation features and microclimate modulation features into the attention submodule. The attention submodule includes a first interactive attention unit and a second interactive attention unit. The first interactive attention unit and the second interactive attention unit have weight parameters that are not shared.

[0070] Step S72: Using the first interactive attention unit, the mosquito density modulation feature is used as the query and the microclimate modulation feature is used as the key and value to perform cross-modal attention interaction and obtain the first interactive attention feature.

[0071] Step S73: Using the second interactive attention unit, cross-modal attention interaction is performed with mosquito density modulation features as the query and microclimate modulation features as the key and value to obtain the second interactive attention features.

[0072] Step S74: Perform mean fusion on the first interactive attention feature and the second interactive attention feature to obtain the attention feature.

[0073] It should be noted that this implementation employs a design of parallel processing by dual interactive attention units followed by mean fusion. The core idea is to capture the conditional response relationship between mosquito density modulation features and microclimate modulation features more comprehensively and robustly through multi-perspective cross-modal interaction. The first and second interactive attention units have the same network structure, both performing cross-modal attention interaction operations with mosquito density modulation features as queries and microclimate modulation features as keys and values, but their weight parameters are not shared. This design allows the two units to learn, from different parameter spaces, the correlation patterns between mosquito density response and the microclimate environment with different emphases during training. For example, one unit may focus more on the immediate impact of temperature changes on the activity density of adult mosquitoes after molting, while the other unit may focus more on the lag effect of humidity accumulation on the larval development and transformation rate. These two units enrich the modeling granularity of cross-modal interaction from different perspectives. Although the interactive attention features extracted from the two units differ in detail, both were obtained under the same breeding source driving conditions and the same input feature space, using the same query-key-value interaction mechanism. Their semantic objects and feature dimensions are completely aligned, and the core semantic they represent is "the conditional contribution pattern of microclimate to mosquito density under breeding source driving conditions." Therefore, mean fusion of the first and second interactive attention features is mathematically and physically feasible. Mean fusion can integrate the complementary information of the two branches, smooth out potential attention weight estimation biases in a single branch, enhance the robustness and generalization ability of the attention features, and ensure that the final fused attention features can more stably reflect the conditional transformation law of "breeding source driving → microclimate regulation → density response" in the complete disaster chain.

[0074] In the above implementation, in order to generate attention features, mosquito density modulation features and microclimate modulation features are loaded into the attention submodule. The attention submodule includes a first interactive attention unit and a second interactive attention unit. The first interactive attention unit and the second interactive attention unit have non-shared weight parameters. Then, through the first interactive attention unit, using the mosquito density modulation features as the query and the microclimate modulation features as the key and value, cross-modal attention interaction is performed to obtain the first interactive attention feature. Then, through the second interactive attention unit, using the mosquito density modulation features as the query and the microclimate modulation features as the key and value, cross-modal attention interaction is performed to obtain the second interactive attention feature. Finally, the first interactive attention feature and the second interactive attention feature are fused by mean to obtain the attention feature.

[0075] As a further implementation of the method, the step of performing cross-modal attention interaction using mosquito density modulation features as the query and microclimate modulation features as the key and value to obtain the first interaction attention feature includes: Step S81: Perform a linear transformation on the mosquito density modulation features to generate the first interactive query matrix.

[0076] Step S82: Perform a linear transformation on the microclimate modulation characteristics to generate the first interaction bond matrix and the first interaction value matrix, respectively.

[0077] Step S83: Calculate the first interaction attention weight matrix based on the first interaction query matrix and the first interaction key matrix.

[0078] Step S84: Based on the first interaction attention weight matrix, the first interaction value matrix is ​​weighted and summed to obtain the first interaction attention feature.

[0079] In the above implementation, in order to generate the first interactive attention feature, the mosquito density modulation feature is linearly transformed to generate the first interactive query matrix, and then the microclimate modulation feature is linearly transformed to generate the first interactive key matrix and the first interactive value matrix, respectively. Then, based on the first interactive query matrix and the first interactive key matrix, the first interactive attention weight matrix is ​​calculated, and then based on the first interactive attention weight matrix, the first interactive value matrix is ​​weighted and summed to obtain the first interactive attention feature.

[0080] This application also discloses a mosquito distribution monitoring and early warning device based on deep learning.

[0081] A mosquito distribution monitoring and early warning device based on deep learning includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements any one of the deep learning-based mosquito distribution monitoring and early warning methods.

[0082] This application also discloses a mosquito distribution monitoring and early warning system based on deep learning.

[0083] refer to Figure 3 A mosquito distribution monitoring and early warning system based on deep learning, comprising: The data acquisition module is used to acquire time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events in each monitoring sub-area within the target agricultural area. Among them, the time-series data on mosquito density is collected by mosquito trapping devices, the time-series data on microclimate is collected by canopy microclimate sensors, and the time-series data on breeding point humidity is collected by water accumulation sensors deployed at irrigation network nodes. The breeding source driving feature generation module is used to take each monitoring sub-area as the target sub-area in turn, and perform semantic mining on the humidity time series data of the breeding points corresponding to the target sub-area based on the global irrigation event record to generate the breeding source driving features of the target sub-area. Among them, the breeding source driving features represent the degree of driving effect of water accumulation in the pipe network nodes caused by irrigation activities on larval hatching and new mosquito emergence. The breeding outbreak risk feature mining module is used to perform deep semantic mining on the microclimate time series data and mosquito density time series data corresponding to the target sub-region based on the breeding source driving features, so as to obtain the breeding outbreak risk features of the target sub-region. Among them, the breeding outbreak risk features represent the response pattern of the microclimate environment driven by the breeding source to the mosquito density outbreak. The early warning module is used to generate a mosquito risk heat map based on the characteristics of breeding and outbreak risks.

[0084] The mosquito distribution monitoring and early warning system based on deep learning of the present invention can implement any of the methods in the mosquito distribution monitoring and early warning method based on deep learning. The specific working process of the mosquito distribution monitoring and early warning system based on deep learning of the present invention can refer to the corresponding process in the above-mentioned mosquito distribution monitoring and early warning method based on deep learning.

[0085] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A mosquito distribution monitoring and early warning method based on deep learning, characterized in that, include: Acquire time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events for each monitoring sub-area within the target agricultural area. The mosquito density time-series data is collected by a mosquito trapping device, the microclimate time-series data is collected by a canopy microclimate sensor, and the breeding point humidity time-series data is collected by water accumulation sensors deployed at irrigation network nodes. Each monitoring sub-area is sequentially taken as the target sub-area. Based on the global irrigation event record, semantic mining is performed on the humidity time series data of the breeding points corresponding to the target sub-area to generate the breeding source driving features of the target sub-area. The breeding source driving features characterize the degree of driving effect of water accumulation in the pipe network nodes caused by irrigation activities on larval hatching and new mosquito emergence. Based on the breeding source-driven characteristics, deep semantic mining is performed on the microclimate time-series data and mosquito density time-series data corresponding to the target sub-region to obtain the breeding outbreak risk characteristics of the target sub-region. The breeding outbreak risk characteristics characterize the response pattern of the microclimate environment driven by breeding sources to mosquito density outbreaks. Based on the aforementioned breeding and outbreak risk characteristics, a mosquito risk heat map is generated.

2. The mosquito distribution monitoring and early warning method based on deep learning according to claim 1, characterized in that, The step of performing semantic mining on the humidity time-series data of the breeding points corresponding to the target sub-region based on the global irrigation event records to generate the breeding source driving features of the target sub-region includes: The humidity time-series data of the breeding points are vectorized to obtain a humidity time-series embedding vector; The global irrigation event records are vectorized to obtain irrigation event embedding vectors; The humidity time-series embedding vector and the irrigation event embedding vector are input into the breeding source semantic mining module, wherein the breeding source semantic mining module includes a first gating unit and a second gating unit, and the first gating unit and the second gating unit have non-shared weight parameters; The first gating unit modulates the humidity temporal embedding vector based on the irrigation event embedding vector to obtain the first irrigation modulation feature. The second gating unit modulates the humidity temporal embedding vector based on the irrigation event embedding vector to obtain the second irrigation modulation feature. The first irrigation modulation feature and the second irrigation modulation feature are fused by mean fusion to obtain the breeding source driving feature of the target sub-region.

3. The mosquito distribution monitoring and early warning method based on deep learning according to claim 2, characterized in that, The step of gating and modulating the humidity temporal embedding vector based on the irrigation event embedding vector to obtain the first irrigation modulation feature includes: A linear mapping is performed on the irrigation event embedding vector to obtain an intermediate mapping vector, wherein the dimension of the intermediate mapping vector is the same as the dimension of the humidity time-series embedding vector; A nonlinear transformation is performed on the intermediate mapping vector to obtain the gated vector; The gate vector is broadcast in the time dimension to generate a gate matrix, wherein the shape of the gate matrix is ​​consistent with that of the humidity time-series embedding vector; The humidity temporal embedding vector is multiplied element-wise with the gating matrix to obtain the first irrigation modulation feature.

4. The mosquito distribution monitoring and early warning method based on deep learning according to claim 1, characterized in that, The step of performing deep semantic mining on the microclimate time-series data and mosquito density time-series data corresponding to the target sub-region based on the breeding source driving characteristics to obtain the breeding outbreak risk characteristics of the target sub-region includes: The microclimate time-series data and the mosquito density time-series data are vectorized to obtain the microclimate embedding vector and the mosquito density embedding vector, respectively. The breeding source driving features, the microclimate embedding vector, and the mosquito density embedding vector are input into the breeding outbreak risk mining module, wherein the breeding outbreak risk mining module includes a breeding source modulation submodule and an attention submodule; Based on the breeding source modulation submodule, using the breeding source driving feature as the query, attention-weighted modulation is applied to the microclimate embedding vector and the mosquito density embedding vector respectively to obtain the microclimate modulation feature and the mosquito density modulation feature. Based on the attention submodule, using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value, cross-modal attention interaction is performed to obtain attention features; The attention features are globally pooled to obtain the breeding outbreak risk features of the target sub-region.

5. The mosquito distribution monitoring and early warning method based on deep learning according to claim 4, characterized in that, The step of using the breeding source modulation submodule as a query to perform attention-weighted modulation on the microclimate embedding vector and the mosquito density embedding vector to obtain microclimate modulation features and mosquito density modulation features includes: The breeding source driving features, the microclimate embedding vector, and the mosquito density embedding vector are loaded into the breeding source modulation submodule, wherein the breeding source modulation submodule includes a first attention modulation unit and a second attention modulation unit; Using the first attention modulation unit, the microclimate embedding vector is subjected to attention-weighted modulation with the breeding source driving feature as the query, to obtain the microclimate modulation feature; The second attention modulation unit uses the breeding source driving feature as a query to perform attention-weighted modulation on the mosquito density embedding vector to obtain the mosquito density modulation feature.

6. The mosquito distribution monitoring and early warning method based on deep learning according to claim 5, characterized in that, The step of using the breeding source-driven features as a query to perform attention-weighted modulation on the microclimate embedding vector to obtain microclimate modulation features includes: A linear transformation is performed on the aforementioned breeding source driving features to generate a modulation query vector; A linear transformation is performed on the microclimate embedding vector to generate a microclimate key vector and a microclimate value vector. The microclimate embedding vector includes embedding vectors for multiple time steps. The microclimate key vector is composed of the key vectors corresponding to each time step, and the microclimate value vector is composed of the value vectors corresponding to each time step. The similarity between the modulation query vector and each of the key vectors is calculated, and the microclimate attention weight distribution is generated by a normalized exponential function. Based on the microclimate attention weight distribution, the value vectors are weighted and summed to obtain the microclimate modulation features.

7. The mosquito distribution monitoring and early warning method based on deep learning according to claim 4, characterized in that, The step of obtaining attention features by performing cross-modal attention interaction based on the attention submodule, using the mosquito density modulation features as the query and the microclimate modulation features as the key and value, includes: The mosquito density modulation feature and the microclimate modulation feature are loaded into the attention submodule, wherein the attention submodule includes a first interactive attention unit and a second interactive attention unit, and the first interactive attention unit and the second interactive attention unit have weight parameters that are not shared. The first interactive attention unit performs cross-modal attention interaction using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value to obtain the first interactive attention feature. The second interactive attention unit performs cross-modal attention interaction using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value to obtain the second interactive attention feature. The first interactive attention feature and the second interactive attention feature are fused by mean fusion to obtain the attention feature.

8. The mosquito distribution monitoring and early warning method based on deep learning according to claim 7, characterized in that, The step of performing cross-modal attention interaction using the mosquito density modulation feature as the query and the microclimate modulation feature as the key and value to obtain the first interaction attention feature includes: A linear transformation is performed on the mosquito density modulation features to generate a first interactive query matrix; A linear transformation is performed on the microclimate modulation features to generate a first interaction bond matrix and a first interaction value matrix, respectively. Based on the first interaction query matrix and the first interaction key matrix, calculate the first interaction attention weight matrix; Based on the first interaction attention weight matrix, the first interaction value matrix is ​​weighted and summed to obtain the first interaction attention feature.

9. A mosquito distribution monitoring and early warning device based on deep learning, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the method of any one of claims 1 to 8.

10. A mosquito distribution monitoring and early warning system based on deep learning, characterized in that, include: The data acquisition module is used to acquire time-series data on mosquito density, microclimate, breeding point humidity, and global irrigation events in each monitoring sub-area within the target agricultural area. The mosquito density time-series data is collected by a mosquito trapping device, the microclimate time-series data is collected by a canopy microclimate sensor, and the breeding point humidity time-series data is collected by water accumulation sensors deployed at irrigation network nodes. The breeding source driving feature generation module is used to take each monitoring sub-area as a target sub-area in sequence, and perform semantic mining on the humidity time series data of the breeding points corresponding to the target sub-area based on the global irrigation event record to generate the breeding source driving feature of the target sub-area. The breeding source driving feature represents the degree of driving effect of water accumulation in the pipe network nodes caused by irrigation activities on larval hatching and new mosquito emergence. The breeding outbreak risk feature mining module is used to perform deep semantic mining on the microclimate time series data and mosquito density time series data corresponding to the target sub-region based on the breeding source driving features, so as to obtain the breeding outbreak risk features of the target sub-region, wherein the breeding outbreak risk features characterize the response pattern of the microclimate environment driven by the breeding source to the mosquito density outbreak; The early warning module is used to generate a mosquito risk heat map based on the breeding and outbreak risk characteristics.

Images